Performance Tuning¶

To keep the installation simple and feature-complete, the default installation script for Kube-OVN does not have performance-specific optimizations. If the applications are sensitive to latency and throughput, administrators can use this document to make specific performance optimizations.

The community will continue to iterate on the performance. Some general performance optimizations have been integrated into the latest version, so it is recommended to use the latest version to get better default performance.

For more on the process and methodology of performance optimization, please watch the video Kube-OVN 容器性能优化之旅。

Benchmarking¶

Because the hardware and software environments vary greatly, the performance test data provided here can only be used as a reference, and the actual test results may differ significantly from the results in this document. It is recommended to compare the performance test results before and after optimization, and the performance comparison between the host network and the container network.

Overlay Performance Comparison before and after Optimization¶

Environment:

• Kubernetes: 1.22.0
• OS: CentOS 7
• Kube-OVN: 1.8.0 Overlay Mode
• CPU: Intel(R) Xeon(R) E-2278G
• Network: 2*10Gbps, xmit_hash_policy=layer3+4

We use qperf -t 60 <server ip> -ub -oo msg_size:1 -vu tcp_lat tcp_bw udp_lat udp_bw to test bandwidth and latency of tcp/udp in 1-byte packets and the host network, respectively.

Overlay， Underlay and Calico Comparison¶

Next, we compare the overlay and underlay performance of the optimized Kube-OVN at different packet sizes with Calico's IPIP Always, IPIP never and the host network.

Environment:

• Kubernetes: 1.22.0
• OS: CentOS 7
• Kube-OVN: 1.8.0
• CPU: AMD EPYC 7402P 24-Core Processor
• Network: Intel Corporation Ethernet Controller XXV710 for 25GbE SFP28

qperf -t 60 <server ip> -ub -oo msg_size:1 -vu tcp_lat tcp_bw udp_lat udp_bw

qperf -t 60 <server ip> -ub -oo msg_size:1K -vu tcp_lat tcp_bw udp_lat udp_bw

qperf -t 60 <server ip> -ub -oo msg_size:4K -vu tcp_lat tcp_bw udp_lat udp_bw

In some cases the container network outperforms the host network, this is because the container network path is optimized to completely bypass netfilter. Due to the existence of kube-proxy, all packets in host network have to go through netfilter, which will lead to more CPU consumption, so that container network in some environments has better performance.

Dataplane performance optimization methods¶

The optimization methods described here are related to the hardware and software environment and the desired functionality, so please carefully understand the prerequisites for optimization before attempting it.

CPU Performance Mode Tuning¶

In some environments the CPU is running in power saving mode, performance in this mode will be unstable and latency will increase significantly, it is recommended to use the CPU's performance mode for more stable performance.

cpupower frequency-set -g performance

NIC Hardware Queue Adjustment¶

In the case of increased traffic, a small buffer queue may lead to significant performance degradation due to a high packet loss rate and needs to be tuned.

Check the current NIC queue length:

# ethtool -g eno1
 Ring parameters for eno1:
 Pre-set maximums:
 RX:             4096
 RX Mini:        0
 RX Jumbo:       0
 TX:             4096
 Current hardware settings:
 RX:             255
 RX Mini:        0
 RX Jumbo:       0
 TX:             255

Increase the queue length to the maximum:

ethtool -G eno1 rx 4096
ethtool -G eno1 tx 4096

Optimize with tuned¶

tuned can use a series of preconfigured profile files to perform system optimizations for a specific scenario.

For latency-first scenarios:

tuned-adm profile network-latency

For throughput-first scenarios:

tuned-adm profile network-throughput

Interrupt Binding¶

We recommend disabling irqbalance and binding NIC interrupts to specific CPUs to avoid performance fluctuations caused by switching between multiple CPUs.

Disable OVN LB¶

The L2 LB implementation of OVN requires calling the kernel's conntrack module and recirculate, resulting in a significant CPU overhead, which is tested to be around 20%. For Overlay networks you can use kube-proxy to complete the service forwarding function for better Pod-to-Pod performance. This can be turned off in kube-ovn-controller args:

command:
- /kube-ovn/start-controller.sh
args:
...
- --enable-lb=false
...

In Underlay mode kube-proxy cannot use iptables or ipvs to control container network traffic, if you want to disable the LB function, you need to confirm whether you do not need the Service function.

FastPath Kernel Module¶

Since the container network and the host network are on different network ns, the packets will pass through the netfilter module several times when they are transmitted across the host, which results in a CPU overhead of nearly 20%. The FastPath module can reduce CPU overhead by bypassing netfilter, since in most cases applications within a container network do not need to use the functionality of the netfilter module.

If you need to use the functions provided by netfilter such as iptables, ipvs, nftables, etc. in the container network, this module will disable the related functions.

Since kernel modules are kernel version dependent, it is not possible to provide a single kernel module artifact that adapts to all kernels. We pre-compiled the FastPath module for part of the kernels, which can be accessed by tunning-package.

You can also compile it manually, see Compiling FastPath Module

After obtaining the kernel module, you can load the FastPath module on each node using insmod kube_ovn_fastpath.ko and verify that the module was loaded successfully using dmesg:

# dmesg
...
[619631.323788] init_module,kube_ovn_fastpath_local_out
[619631.323798] init_module,kube_ovn_fastpath_post_routing
[619631.323800] init_module,kube_ovn_fastpath_pre_routing
[619631.323801] init_module,kube_ovn_fastpath_local_in
...

OVS Kernel Module Optimization¶

OVS flow processing including hashing, matching, etc. consumes about 10% of the CPU resources. Some instruction sets on modern x86 CPUs such as popcnt and sse4.2 can speed up the computation process, but the kernel is not compiled with these options enabled. It has been tested that the CPU consumption of flow-related operations is reduced to about 5% when the corresponding instruction set optimizations are enabled.

Similar to the compilation of the FastPath module, it is not possible to provide a single kernel module artifact for all kernels. Users need to compile manually or go to tunning-package to see if a compiled package is available for download.

Before using this kernel module, please check if the CPU supports the following instruction set:

cat /proc/cpuinfo  | grep popcnt
cat /proc/cpuinfo  | grep sse4_2

Compile and Install in CentOS¶

Install the relevant compilation dependencies and kernel headers:

yum install -y gcc kernel-devel-$(uname -r) python3 autoconf automake libtool rpm-build openssl-devel

Compile the OVS kernel module and generate the corresponding RPM:

git clone -b branch-2.17 --depth=1 https://github.com/openvswitch/ovs.git
cd ovs
curl -s  https://github.com/kubeovn/ovs/commit/2d2c83c26d4217446918f39d5cd5838e9ac27b32.patch |  git apply
./boot.sh
./configure --with-linux=/lib/modules/$(uname -r)/build CFLAGS="-g -O2 -mpopcnt -msse4.2"
make rpm-fedora-kmod
cd rpm/rpmbuild/RPMS/x86_64/

Copy the RPM to each node and install:

rpm -i openvswitch-kmod-2.15.2-1.el7.x86_64.rpm

If you have previously started Kube-OVN and the older version of the OVS module has been loaded into the kernel. It is recommended to reboot the machine to reload the new version of the kernel module.

Compile and Install in Ubuntu¶

Install the relevant compilation dependencies and kernel headers:

apt install -y autoconf automake libtool gcc build-essential libssl-dev

Compile the OVS kernel module and install:

apt install -y autoconf automake libtool gcc build-essential libssl-dev

git clone -b branch-2.17 --depth=1 https://github.com/openvswitch/ovs.git
cd ovs
curl -s  https://github.com/kubeovn/ovs/commit/2d2c83c26d4217446918f39d5cd5838e9ac27b32.patch |  git apply
./boot.sh
./configure --prefix=/usr/ --localstatedir=/var --enable-ssl --with-linux=/lib/modules/$(uname -r)/build
make -j `nproc`
make install
make modules_install

cat > /etc/depmod.d/openvswitch.conf << EOF
override openvswitch * extra
override vport-* * extra
EOF

depmod -a
cp debian/openvswitch-switch.init /etc/init.d/openvswitch-switch
/etc/init.d/openvswitch-switch force-reload-kmod

If you have previously started Kube-OVN and the older version of the OVS module has been loaded into the kernel. It is recommended to reboot the machine to reload the new version of the kernel module.

Using STT Type Tunnel¶

Common tunnel encapsulation protocols such as Geneve and Vxlan use the UDP protocol to encapsulate packets and are well supported in the kernel. However, when TCP packets are encapsulated using UDP, the optimization and offload features of modern operating systems and network cards for the TCP protocol do not work well, resulting in a significant drop in TCP throughput. In some virtualization scenarios, due to CPU limitations, TCP packet throughput may even be a tenth of that of the host network.

STT provides an innovative tunneling protocol that uses TCP formatted header for encapsulation. This encapsulation only emulates the TCP protocol header format without actually establishing a TCP connection, but can take full advantage of the TCP optimization capabilities of modern operating systems and network cards. In our tests TCP packet throughput can be improved several times, reaching performance levels close to those of the host network.

The STT tunnel is not pre-installed in the kernel and needs to be installed by compiling the OVS kernel module, which can be found in the previous section.

Enable STT tunnel:

kubectl set env daemonset/ovs-ovn -n kube-system TUNNEL_TYPE=stt

kubectl delete pod -n kube-system -lapp=ovs

微信群 Slack Twitter Support

Comments